How much electricity home AI uses in a year

Three states, three numbers

State	Draw of the whole machine
Waiting for a question	120 W
Typical work	256 W
Heaviest configuration	352 W
Momentary peak	580 W

The spikes to 580 W last seconds and fall on the most demanding stretches of computation. We chose the power supply with headroom — by design, the machine can reach about 670 W.

The bill for a year

All the costs below use our Polish tariff of €0.26 ($0.29) per kilowatt-hour — plug in your local rate to adjust them:

Scenario	Per year	Cost
On around the clock, no work	1051 kWh	€270 ($310)
Around the clock, 2 h of heavy work daily	1183 kWh	€300 ($350)
On for 12 h, including 2 h of work	657 kWh	€170 ($195)

The proportions say it all: adding two hours of real work per day raises the bill by less than 12%. The rest is idling. If you want to lower the cost of home AI, do not cut back on using it — cut back on the hours the machine sits switched on doing nothing.

Why idling costs this much

The cards alone draw about 52 W at rest — close to half the whole computer’s consumption. The cards drop then to their lowest power-saving state: core at 210 MHz, fans stopped, bus throttled. That is the floor, and we verified there is no way below it:

Closing programs changes nothing. The draw does not shift even half a watt after stopping the model server and the graphical environment. The driver maintains this state regardless of whether anyone is using the cards.

Unloading the driver makes things 70 W worse. Without the driver the cards sit at factory settings, with no power management — the plug then shows 201 W instead of 130. “Unloading the driver while idle” is a move in the wrong direction.

Deeper card sleep is unavailable on this motherboard. The cards’ hardware could do it, but the motherboard does not expose the required mechanism. On a newer platform this point would look different — worth checking if you are building a machine from scratch.

A curiosity we cannot explain: at identical settings, one card consistently draws 9 W more than the other (21 versus 30 W). Repeatable, checked many times.

What real work costs

The most tangible test: forty runs in which the model wrote multi-module programming projects with tests and documentation, all on its own. Seventeen hours of continuous load.

The total: 5.166 kWh, that is €1.32 ($1.52). About €0.03 ($0.04) per project. Of that amount, 40% went into simply keeping the computer on — not into computing.

A check on the measurement: the plug’s meter showed 5.21 kWh against 5.17 kWh calculated by integrating instantaneous power. A discrepancy of 0.8%.

Which model is more frugal — and why the answer is not obvious

Configuration	Energy	Average power	Peak
Sparse, with reasoning	0.955 kWh	256 W	444 W
Sparse, without reasoning	1.023 kWh	274 W	498 W
Dense, without reasoning	1.288 kWh	350 W	580 W
Dense, with reasoning	1.618 kWh	352 W	569 W

The spread between the extreme configurations is 1.7× — on the same set of tasks and in practically the same time.

The dense model draws about 30% more power despite identical working time, because it keeps the cards busy 56–58% of the time, while the sparse one does so 35–39%. The latter works in bursts, interleaved with running tests and tools — and in those pauses the cards rest.

The most interesting result concerns reasoning, the mode in which the model first deliberates and then answers. On the sparse model it saves electricity (−7% energy, the lowest power peak in the whole table), because better decisions mean fewer failed attempts. On the dense model it costs +26% energy — the single largest overhead in the whole test. The same feature, two opposite effects, depending on the model’s architecture.

A point of reference

A bill in the range of €170–300 ($195–350) a year places such a machine roughly where an older refrigerator or a marine aquarium stands. It is not an invisible cost, but neither is it a cost that decides whether the undertaking makes sense — compared with subscriptions to cloud services it usually comes out favorably, and the data stays at home.

If you are looking for a single lever: schedule the machine to switch off at night. That is a difference on the order of €95–115 ($105–135) a year — more than anything you can gain from the choice of model.

---

Measured with a smart plug, instantaneous power integrated over the time windows of the runs. A methodological note: the plug measures the whole computer, so the numbers also include the work of tools launched by the model.